Publication: Performance characterization of bag-valve-mask (BVM) compression using machine learning
Loading...
Date
2024-02
Authors
Sanjivan Muthu Kumar
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Medical staff face issues when ventilating patients manually using the Bag-Valve-Mask (BVM) for long periods to resuscitate patients unable to breathe properly on their own. As for ICU mechanical ventilators in hospitals, medical specialists must check on patients frequently and adjust settings manually. Currently, there are portable ventilators available in the market that aid in supplying oxygen to patients, however the usage of ML is rare, and they do not take into account various variables which are deemed important in patient recovery. In this research, the BVM was used to perform ventilation using manual and automated methods, after which machine learning (ML) study was done. The first objective was to predict the average tidal volume using artificial neural network (ANN) and boosted decision tree regression algorithms. The R2 value obtained from manual ventilation using ANN was 0.738861, whereas the boosted decision tree model scored 0.600049. Thus, ANN was used on the automated ventilation system to compare its performance with the manual, where an R2 value of 0.978604 was obtained after removing unwanted features. When compared with the manual model, a 32% increase in R2 was obtained. K-fold cross validation was carried out to test the manual and automated models in a bigger data space, where the standard deviation of the automated model was significantly lower, indicating lower variability within its dataset. The outcome of the study suggests that the automated system predicts the experiment data better than the manual system when utilizing ANN. Another objective of this research included conducting ML study using data collected from an ICU mechanical ventilator to provide a setting recommendation for a particular patient using linear and Poisson regression, where linear regression scored a R2 value of 0.936, whereas the Poisson model scored 0.836 when tested on tidal volume (TV) setting. Thus, linear regression was used to perform ML on the TV setting, fraction of inspired oxygen (FiO2) setting, and positive end-expiratory pressure (PEEP) setting, where TV setting scored the highest R2 values overall. To validate the TV setting formula obtained through Microsoft (MS) Azure, three experiments were conducted using a ventilator prototype on an artificial test lung for validation. The experiments yielded error results ranging from 53% to 79%, indicating that the TV setting values obtained from the prototype were incomparable to mechanical ventilator data. Extensive research is needed to compare the results between BVM ventilators and ICU mechanical ventilators.